strategies for containing an emerging influenza pandemic ... · 1 strategies for containing an...
TRANSCRIPT
Title Strategies for containing an emerging influenza pandemic inSoutheast Asia
Author(s) Ferguson, NM; Cummings, DAT; Cauchemez, S; Fraser, C; Riley,S; Meeyai, A; Iamsirithaworn, S; Burke, DS
Citation Nature, 2005, v. 437 n. 7056, p. 209-214
Issued Date 2005
URL http://hdl.handle.net/10722/54275
Rights Creative Commons: Attribution 3.0 Hong Kong License
1
Strategies for containing an emerging influenza pandemic in SE Asia.
Neil M. Ferguson1, Derek Cummings2 Simon Cauchemez3, Christophe Fraser1, Steven
Riley4, Aronrag Meeyai1, Sopon Iamsirithaworn5 & Donald Burke2
1Department of Infectious Disease Epidemiology, Faculty of Medicine, Imperial College
London, St Mary's Campus, Norfolk Place, London W2 1PG, UK
2 Department of International Health, Johns Hopkins Bloomberg School of Public Health,
615 North Wolfe Street, Baltimore, Maryland, 21205, USA
3 INSERM U707, 27 rue Chaligny, Paris, 75571 cedex 12, France
4 Department of Community Medicine, 5/F William M.W. Mong Block, Faculty of Medicine
Building, 21 Sassoon Road, Hong Kong
5Bureau of Epidemiology, Department of Diseases Control, Ministry of Public Health,
Tivanonda Road, Nonthaburi 11000, Thailand
Keywords: influenza, pandemic, H5N1, mathematical model, antivirals, neuraminidase
inhibitor
2
Highly pathogenic H5N1 influenza A viruses are now endemic in avian populations in
Southeast (SE) Asia, and human cases continue to accumulate. While currently
incapable of sustained human-to-human transmission, H5N1 represents a serious
pandemic threat due to the risk of a mutation or reassortment generating a virus with
increased transmissibility. Identifying public health interventions which may be capable
of halting a pandemic in its earliest stages is therefore a priority. Here we use a
simulation model of influenza transmission in SE Asia to evaluate the potential
effectiveness of targeted mass prophylactic use of antiviral drugs as a containment
strategy. Other interventions aimed at reducing population contact rates are also
examined as reinforcements to an antiviral-based containment policy. We demonstrate
that elimination of a nascent pandemic may be feasible using a combination of
geographically targeted prophylaxis and social distancing measures if the basic
reproduction number of the novel virus is below 1.8. We predict that a 3 million course
stockpile should be sufficient for elimination. Policy effectiveness critically depends on
the ascertainment rate of clinical cases and the speed with which antiviral drugs can be
distributed.
The continuing spread of the H5N1 highly pathogenic avian influenza (HPAI) in wild and
domestic poultry in SE Asia represents the most serious human pandemic influenza risk for
decades1,2. Great potential benefits would be gained from any intervention capable of
containing the spread of a pandemic strain and eliminating it from the human population.
However, the rapid rate of spread of influenza – as witnessed both in annual epidemics and
past pandemics3-5– poses a significant challenge to the design of a realistic control strategy.
The basic reproduction number6, 0R , quantifies the transmissibility of any pathogen, being
defined as the average number of secondary cases generated by a typical primary case in an
entirely susceptible population. If 0 1R > a disease can spread, while if 0 1R < , chains of
transmission will inevitably die out. Hence the goal of control policies is to reduce 0R to
below 1, by eliminating a proportion 01 1/ R− of transmission. This can be achieved in three
ways: (a) by reducing contact rates in the population (through ‘social distance measures’); (b)
3
by reducing the infectiousness of infected individuals (through treatment or isolation); or (c)
by reducing the susceptibility of uninfected individuals (by vaccination or antiviral
prophylaxis).
Vaccination and antiviral drugs offer protection against infection and clinical disease.
However while effective vaccines exist for interpandemic flu, candidate H5N1 vaccines have
unproven effectiveness7 and production delays would in any case limit availability in the first
months of a pandemic. Antiviral agents – and in particular the neuraminidase inhibitors,
which show experimental effectiveness against all influenza A subtypes8,9 – are therefore a
key plank of recently revised pandemic preparedness plans in several countries10.
For antivirals to significantly reduce transmission, prophylactic use is necessary. Large-scale
prophylaxis has the potential to limit spread substantially in a developed country context11,
but the very large stocks of drug necessary make this policy impractical if the pandemic is
already global. However, might such a policy nonetheless be a feasible strategy if applied at
the source of a new pandemic, when repeated human-to-human transmission is first
observed? We address this question here, focusing on identification of the threshold level of
transmissibility below which containment of any new pandemic strain might be feasible.
Modelling pandemic spread
We model spread in SE Asia, as that region remains the focus of the ongoing avian H5N1
epidemic and is where most human cases have occurred. Greater data availability led us to
model Thailand, not a perceived greater risk of emergence compared with other countries in
the region; we believe our conclusions are also valid for other parts of SE Asia.
We constructed a spatially-explicit simulation of the 85 million people residing within
Thailand and in a 100km zone of contiguous neighbouring countries. The model explicitly
incorporates households, schools and workplaces as these are known to be the primary
contexts of influenza transmission12-14(see Fig1 and Methods) and because control measures
4
can readily target these locations. Random contacts in the community associated with day-to-
day movement and travel are also modelled..
Natural history and transmission parameters
Fundamental to the feasibility of any containment strategy is the transmissibility of the
emergent virus, as quantified by 0R . Reliable past estimates of transmissibility are rare,
perhaps due to the antigenic diversity of the influenza and consequent complex impact of
population immunity on transmission.
We reanalysed incubation period and household transmission data (see Methods) and derived
new natural history parameters which predict a profile of infectiousness through time which
is remarkably consistent with viral shedding data from experimental infection studies (see Figure 1g15. This profile gives an estimate of the serial interval or generation time, gT (the
average interval from infection of one individual to when their contacts are infected), of 2.6
days – to be compared with the value of ~4 days assumed by most past modelling16. Reanalysis of both US and UK 1918 mortality data using this new value of gT revises
pandemic influenza 0R estimates5 downwards to approximately 1.8 (Figure 1a) – yielding a
predicted infection attack rate during a pandemic of 50-60%, consistent with what was seen
in the 1st and 2nd waves of past pandemics (see Figure 2b). A value of 1.8 is also consistent
with annual interpandemic attack rates seen for households where all members were highly
susceptible to the prevalent strain17 (see Supplementary Information). We additionally
assume that 50% of infections result in clinically recognisable symptoms, with the other 50%
being too mild to be diagnosed clinically 18.
It is not certain that these parameter estimates would be applicable to any new pandemic
strain. It is possible that the mutations or reassortment event that give rise to the new viral
strain might initially only increase its transmissibility a little over the 0R =1 threshold for self-
sustaining transmission. In that case, additional mutations would need to accumulate for viral
fitness to increase to its maximum. Given the extended viral shedding (and symptomatic
5
disease) seen in severe human cases of avian H5N1 infection, this also might mean that the
gT of the initial pandemic strain could be considerably greater than for currently circulating
human influenza viruses. We therefore examine the ability of control measures to contain
pandemic spread not just at a single value of 0R , but for different values in the range
1< 0R <2, and analyse model sensitivity to the assumed value of gT .
Baseline epidemic dynamics.
We consider the scenario that a novel transmissible ( 0R >1) pandemic strain arises as a result
of mutations or a reassortment event in a single individual infected with an avian virus. We
seed simulations with a single infection in the most rural (i.e. lowest population density) third
of the population, assuming rural populations are most likely to be exposed to the avian virus.
Figure 2 shows the typical pattern of spread for an emergent pandemic initiated by such a
seeding event assuming 0R =1.5 – though note that for low 0R most epidemics seeded by a
single individual go extinct by chance before becoming established in the population.
The pattern of spatial spread (Figure 2a and Supplementary Information video 1) is of
interest: for the first 30 days, cases tend to be limited to the region around the seeding
location, with few ‘sparks’ outside that area. However, as case numbers increase
exponentially, so does the frequency with which infection events span large distances, and
the epidemic rapidly transforms from being predominantly local to country-wide between day
60 and 90. Any containment policy needs to be effective before this transition – in part
because logistic constraints are likely to preclude containment of a widely disseminated
epidemic, but also because the probability of international export of infection becomes high
once case numbers reach the thousands19.
For 0R =1.5, the epidemic in the modelled population of 85 million peaks around day 150,
and is largely over by day 200, at which point 33% of the population have been infected. At
0R =1.8 the epidemic peaks around day 100 and infects around 50% of the population.
6
Impact of antiviral prophylaxis
In containment strategies, we focus on 2 principal outcome measures: (a) the probability of
preventing a large outbreak (which would eventually lead to a global pandemic); (b) the
number of courses of drug (here assumed to be oseltamivir) required to achieve containment.
Blanket prophylaxis of an entire country or region should be able to eliminate a pandemic
virus with an 0R of 3.6 or greater (see Methods). However, such a policy would require
enough drug to prophylax everyone for up to 3 weeks (i.e. at least 2 courses per person), and
hence is unfeasible . Targeted strategies are therefore needed which minimise drug usage
while maximising impact.
Social targeting is the most straightforward approach – namely prophylaxing individuals in
the same household, school or workplace as a newly diagnosed symptomatic case.
Unfortunately, if such a policy is only initiated after 20 or more cases, purely social targeting
only has a has ≥90% probability of eliminating the pandemic strain if 0R ≤1. 25 (lowest curve
of Figure 3a; see also Supplementary Information). In reality, at least 10 cases might need to
be detected to be sure that viral transmissibility had significantly increased20, and detection
and decision-making delays could easily mean 20-30 cases had arisen before policy initiation.
A containment policy will therefore probably need to go beyond social targeting to succeed.
Since most community contacts are local, geographic targeting – namely when a case is
detected, prophylaxing the whole population in neighbourhood of the household of the case –
is an obvious policy extension, though one which undoubtedly will dramatically increase the
logistical challenges to delivery. In the absence of detailed administrative boundary data, we
simulated geographic targeting as the prophylaxis of the population within a ring of a certain
radius centred around each detected case, though in reality targeting administrative areas is
likely to be more practical. For socially or geographically prophylaxis, we assume individuals
are given a single course of 10 days of drug, after which time they come off drug unless more
cases have arisen in their vicinity, in which case a second round of prophylaxis is delivered.
The policies therefore automatically cease within 10 days of the last case being reported.
7
Our analysis indicates that the substantial additional effort required to deliver a geographic
policy pays substantial dividends in policy effectiveness. With a 2 day delay from case onset
to prophylaxis, a 5km ring policy is capable of containing pandemics with an 0R of 1.5
(Figure 3a) at the cost of an average of 2 million courses (Figure 3b) - though the maximum
number of courses needed can by an (unfeasible) order of magnitude for scenarios where
cases arise in Bangkok at an early stage of the outbreak. Policy effectiveness increases with
the radius of the treatment ring selected (though little benefit is gained from exceeding
10km), but so does the number of courses required (Figure 3b). Policy outcome is still
sensitive to the speed of case detection and drug delivery, but containment is always
substantially better than for the purely socially targeted policy (Figure 3d).
Since pure radial prophylaxis is costly in terms of drug, we also examined a policy variant
which limits the number of people targeted for prophylaxis per case by only targeting the
nearest m people (where m=10,000-50,000) within 10km of a newly diagnosed case. In areas
of low population density, this drug-sparing policy has the impact of a pure 10km policy, but
in high density areas many fewer courses of drug are used. The improved effectiveness in
rural areas outweighs decreased effectiveness in urban areas resulting in greater impact than a
pure 5km policy and much lower drug use (Figure 3e-f).
Epidemiologically, elimination either occurs because the treatment strategy reduces 0R to
below 1, or because it reduces it to close to 1 when the epidemic is small, hence enhancing
the probability of random extinction. In scenarios where the pandemic strain is successfully
eliminated, geographic spread is usually limited. For example, the root-mean-square (rms)
radius of spread is 27km for 0R =1.5 using the 5km radial geographic targeting strategy.
When containment is successful, total case numbers are also limited to an average of fewer
than 150 cases.
Policies to increase social distance
8
Measures to increase social distance have been employed in past pandemics, and remain
important options for responding to future pandemics1. Predicting the impact of policies such
as closing schools and workplaces is difficult, however, as potentially infectious contacts may
be displaced into other settings. Furthermore, it is likely that population contact rates change
spontaneously (as well as a result of policy) during severe epidemics (e.g. 1918) in response
to the perceived risk. Therefore the estimates of pandemic transmissibility we derive from
past pandemics may implicitly incorporate the effects of some degree of social distancing.
We are therefore deliberately conservative in the assumptions made as to the impact of school
and workplace closure here, by assuming household and random contact rates increase by
100% and 50% respectively for individuals no longer able to attend school or work. Figure 4a
illustrates how adding area-based school and workplace closure, to a drug-sparing
prophylaxis policy increases policy effectiveness significantly, with the combined policy
having a >90% chance of elimination for 0R =1.7.
Quarantine zones – in which movements in and out of the affected area are restricted – are
another strategy for enhancing containment, and may in any case be thought necessary to
prevent population flight from affected areas or people deliberately entering prophylaxis
zones to receive drug. Figure 4a (see also Supplementary Information video 2) shows that
such a strategy can dramatically boost the effectiveness (to 90% containment at 0R =1.8) of a
radial geographic targeted prophylaxis even if only 80% effective at reducing movements.
Combining school and workplace closure with area quarantine and prophylaxis further
increases policy effectiveness (90% containment at 0R =1.9) and as importantly, the
robustness of the policy to shortcomings in case identification or treatment rates. For all these
policies, containment is typically achieved after fewer then 200 cases have been detected.
Logistical constraints and sensitivity to parameter assumptions
Other constraints may affect the ability of public health authorities to deliver containment
policies. Figures 4c shows that size of antiviral stockpile can have a substantial effect on
9
policies which use pure radial geographic prophylaxis since very large numbers of courses
are required to prophylax around cases arising in large urban areas. However policies
employing drug-sparing geographically-targeted prophylaxis (Figure 4d) retain high
effectiveness so long as at least 3 million drug courses are available. For the scenarios where
containment fails given a finite stockpile, Figure 4e shows that even an unsuccessful
containment strategy can delay wide-scale spread by a month or more – a potentially critical
window of opportunity for accelerating vaccine production.
Another possible constraint is that capacity to implement these containment policies may not
be present in all countries in the region, A policy restricted to one country alone may have a
substantially reduced chance of success (Figure 4f and Supplementary Information video 3),
should the initial case cluster arise in a border region.
Multiple assumptions inevitably need to be made in undertaking preparedness modelling for a
future emergent infection. Sensitivity analyses are therefore critical to assessing the
robustness of policy conclusions. Here, critical assumptions not already discussed include (a) the ratio of within-place to community transmission, (b) the expected generation time, gT , of
a new pandemic strain (largely determined by the duration of viral shedding and therefore
infectiousness), (c) the level of heterogeneity in individual infectiousness (e.g.
‘superspreaders’21), (d) antiviral efficacy/take-up, and (e) the sensitivity and specificity of
case detection during the control programme. The impact of these assumptions on model
output is presented in the Supplementary Information, but in summary, (d) and (e) are – as
one might expect – the most critical. If antiviral coverage or efficacy is considerably less than
assumed, then policy effectiveness is substantially reduced. Similarly, if surveillance picks up
fewer than 40% of infections (i.e. 80% of symptomatic cases), again policy effectiveness is
reduced. Poor surveillance specificity (i.e. false positives) has an indirect effect on
effectiveness through wasted drug and logistical capacity..
Conclusions
10
We have shown that containment and elimination of an emergent pandemic strain of
influenza at the point of origin is potentially feasible using a combination of antiviral
prophylaxis and social distance measures. A key conclusion is the need for multiple
approaches: simple socially-targeted prophylaxis is unlikely to be sufficient if the emergent
virus has transmissibility near that of past pandemic viruses. Geographically targeted policies
are needed to achieve high levels of containment, with area quarantine being particularly
effective at further boosting policy effectiveness. The only scenario under which purely
socially targeted strategies might be sufficient would be if viral transmissibility evolved
incrementally and the emergent virus initially had 0R only slightly above one (see
Supplementary Information); however 0R will be probably be uncertain at the time
containment policies have to be implemented, arguing for policies be precautionary in
assuming transmissibility will be comparable with that seen in past pandemics.
A number of key criteria need to be met for a high probability of success: (a) rapid
identification of the original case cluster, (b) rapid and sensitive case detection and then
delivery of treatment to targeted groups – preferably within 48 hours of a case arising, (c)
effective delivery of treatment to a high proportion of the targeted population – preferably
>90%, (d) sufficient stockpiles of drug – preferably 3+ million courses of oseltamivir, (e)
population cooperation with the containment strategy and, in particular, any social distance
measures introduced, (f) international cooperation in policy development, epidemic
surveillance, and control strategy implementation. Lastly, containment is unlikely if 0R
exceeds 1.8 for the new pandemic strain. While our analysis of past pandemics suggests that
transmissibility will be in this range, it is unlikely that sufficient data will exist to verify this
before a containment policy needs to be introduced.
The mathematical model we have used to examine the feasibility of pandemic containment is
perhaps the largest-scale detailed epidemic microsimulation yet developed. A key modelling
goal was parsimony. While the representation of the population is detailed, this detail is
underpinned by available demographic data. The natural history parameters used here have
been estimated from primary data on existing influenza strains. The model has 5 key
11
transmission parameters, of which 2 were estimated from household data and the remaining 3
were qualitatively calibrated to historical age-dependent attack rates. We believe that this
type of simulation will increasingly become a standard tool for preparedness planning and
modelling of novel disease outbreaks.
Given the set of criteria listed above for successful containment, the obstacles to practical
implementation of such a strategy are undoubtedly formidable. Surveillance is perhaps the
single greatest challenge. Success depends on early identification of the first cluster of cases
caused by the pandemic strain20, and on detection of a high proportion of ongoing cases.
Some level of mildly symptomatic infection is to be expected (and has been observed for
human H5N1 infections22), but key to successful containment is the proportion of such cases
and their infectiousness. Should the high pathogenicity of recently reported human infections
with the H5N1 virus be even partly maintained, then containment might paradoxically be
more likely as case-ascertainment levels would be higher.
Achieving the rapid delivery of antiviral drug to a large proportion of the population poses
many challenges. Thailand, the country modelled here, is one of the best prepared and
equipped countries in the region in being able to implement a large-scale and very rapid
public health intervention. Other countries in the region need considerable development input
in basic healthcare and disease surveillance infrastructure in order to meet the needs of
containment.
Antiviral resistance represents an as yet unquantifiable challenge to a prophylaxis-based
containment strategy. Key is not whether genotypic or clinical resistance is seen in a
percentage of individuals, but whether resistant viruses are capable of self-sustaining
transmission (i.e. have 0R >1). As yet, the evidence is indicative that fitness deficits mean
transmissibility is limited for oseltamivir-resistant strains23,24 but the possibility that
compensatory mutations which increase transmissibility might be selected cannot be ruled
out completely. If a transmissible resistant strain did emerge during a containment policy, it
would be essential for prophylaxis to cease, lest wild-type virus be eliminated and the world
be left with a pandemic of resistant virus. If prophylaxis were abandoned, the likely higher
12
fitness of wild type would give every chance that the resistant strain would then be excluded
from the population.
A feasible strategy for containment of the next pandemic offers the potential of preventing
millions of deaths. It is therefore in the interest of all countries to contribute to ensuring the
resources, infrastructure and collaborative relationships are in place within the region
currently most likely to be the source of a new pandemic. Even if the challenges are great, the
costs of failure are potentially so catastrophic that it is imperative for the international
community to prepare now to ensure containment is given the best possible chance of
success.
Methods
Demographic data: The model used Landscan data25 to generate a simulated population
realistically distributed across geographic space (Figure 1a). Thai census data26,27 on
household size and age distributions was used for demographic parameterisation (Figure
1b,c). Thai National Statistical Office data27 was used to determine the number and
proportions of children in school as a function of age, and Thai Dept. of Education28 data on
24,000 schools was used to determine the distribution of school sizes (Figure 1d). Data on
travel distances within Thailand was limited; here we used data collected in the 1994
National Migration Survey29,30 on distances travelled to work (Figure 1e and see
Supplementary Information) to estimate movement kernel parameters. The best fit kernel had asymptotic power law form as a function of distance d given by ( ) ( )~ 1 1 / bf d d a⎡ ⎤+⎣ ⎦ ,
where a=4 km and b=3.8. Thai workplace sizes31 also follow a power law distribution32 with
an estimated maximum single workplace size of approximately 2300, and a mean of 21.
Disease data: The natural history of any H5 based pandemic strain will not be known until it
emerges, so we used parameter estimates for current human influenza subtypes, and
employed sensitivity analyses to investigate what impact deviation from these estimates
would have on policy effectiveness (see Supplementary Information). The mean and
13
standard deviation of the incubation period distribution were estimated as 1.48 and 0.47 days
respectively from data on a multiple exposure event occurring on an aeroplane33
We adopt a more biologically realistic approach than most past modelling studies (though see
ref. 34), and rather than assuming infectiousness is constant from the end of the latent period to recovery, we model it as a function, ( )Tκ (assumed normalized), depending on the time
elapsed from the end of the latent period. The generation time, gT , is just given by the mean
latent period plus ( )0
T T dTκ∞
∫ . Experimental infection data35 indicates the start of
symptoms to be coincident with a sharp increase in viral shedding, so we assume
infectiousness starts at the end of the incubation period. We further assume a 0.25 day delay
from when symptoms start to when diagnosis or health-care seeking behaviour is likely. We employed Bayesian methods (see Supplementary Information) to estimate ( )Tκ from data
collected in a recent household study of respiratory disease incidence36,37. Combined with the
estimated incubation period distribution, this gives the profile of infectiousness shown in Figure 1f. gT is estimates as 2.6 days (95% credible interval: 2.1-3.0) – shorter than
previously assumed (though see ref. 38).
Transmission model: The model is a stochastic, spatially structured, individual-based
discrete time simulation. Individuals are co-located in households, with households being
constructed to reflect typical generational structure while matching empirical distributions of
age structure and household size for Thailand (Figure 1b,c). Households are randomly
distributed in the modelled geographic region with a local density determined by the
Landscan data25. In any timestep of T∆ =0.25 days, a susceptible individual i has probability
[ ]1 exp i Tλ− − ∆ of being infected, where iλ is the instantaneous infection risk for individual i.
Infection risk comes from 3 sources: (a) household, (b) place, and (c) random contacts in the
community. The last of these depends on distance and represents random contacts associated
with movements and travel, and is the only means by which infection can cross national
borders. Analysis of household infection data (see Supplementary Information), gave a
within-household 0R of 0.6 and an overall 0R of 1.8. We partition non-household
transmission to give levels of within-place transmission comparable with household
14
transmission (i.e. 0 0.6R ≈ ) and to qualitatively match 1957 pandemic age-specific attack
rates. When varying 0R , the relative proportions of household, place and community
transmission were kept fixed. Full model details are given in the Supplementary Information.
Antiviral drug action: We use recent statistically rigorous estimates of antiviral efficacy39,
but these are broadly consistent with previous estimates23. Prophylaxis of uninfected
individuals is assumed to reduce susceptibility to infection by 30%, infectiousness if infection
occurs by 60%, and the probability of clinically recognisable symptoms by 65%39. In theory
therefore, blanket prophylaxis of a population should be able to contain a pandemic with an
0R of [ ]1/ (1 0.6)(1 0.3)− − , approximately 3.6. Treatment of a symptomatic case is assumed
to reduce infectiousness by 60% from when treatment is initiated. Overall, for the parameter
values used here, antiviral treatment of a symptomatic case can reduce total infectiousness
throughout the course of infection by a maximum of 28%.
Correspondence and requests for materials should be addressed to N.M.F.
(email: [email protected]).
Acknowledgements We thank the National Institute of General Medical Sciences MIDAS Program (U01 GM070708-01, NMF, DC & DB), the Medical Research Council (NMF), the Royal Society (NMF & CF), the Howard Hughes Medical Institute (NMF), the Research Fund for the Control of Infectious Diseases of the Hong Kong SAR government (SR) and INSERM (SC) for research funding. We thank Fabrice Carrat for providing household data used in this study, and Nancy Cox, Fred Hayden, Ben Schwartz, Klaus Stohr and members of the MIDAS consortium for useful discussions. We thank the MIDAS informatics group for computational resources.
Supplementary information accompanies this paper on the Nature website.
15
References 1. World Health Organisation. Avian influenza: assessing the pandemic threat.
http://www.who.int/csr/disease/influenza/en/H5N1-9reduit.pdf (2005). 2. Abbott, A. & Pearson, H. Fear of human pandemic grows as bird flu sweeps through
Asia. in Nature 472-473 (2004). 3. Spicer, C. C. & Lawrence, C. J. Epidemic influenza in Greater London. Journal of
Hygiene 93, 105-12 (1984). 4. Fleming, D. M., Zambon, M., Bartelds, A. I. M. & Jong de, J. C. The duration and
magnitude of influenza epidemics: A study of surveillance data from sentinel general practices in England, Wales and the Netherlands. Europeean Journal of Epidemiology 15, 467-473 (1999).
5. Mills, C. E., Robins, J. M. & Lipsitch, M. Transmissibility of 1918 pandemic influenza. Nature 432, 904-6 (2004).
6. Anderson, R. M. & May, R. M. Infectious Diseases of Humans; Dynamics and Control (Oxford University Press, Oxford, 1991).
7. Stohr, K. & Esveld, M. Will Vaccines Be Available for the Next Influenza Pandemic? Science 306, 2195-2196 (2004).
8. Leneva, I. A., Roberts, N., Govorkova, E. A., Goloubeva, O. G. & Webster, R. G. The neuraminidase inhibitor GS4104 (oseltamivir phosphate) is efficacious against A/Hong Kong/156/97 (H5N1) and A/Hong Kong/1074/99 (H9N2) influenza viruses. Antiviral Res 48, 101-15 (2000).
9. Brooks, M. J., Sasadeusz, J. J. & Tannock, G. A. Antiviral chemotherapeutic agents against respiratory viruses: where are we now and what's in the pipeline? Curr Opin Pulm Med 10, 197-203 (2004).
10. Coombes, R. UK stocks up on antiviral drug to tackle flu outbreak. Bmj 330, 495 (2005).
11. Longini, I. M., Jr., Halloran, M. E., Nizam, A. & Yang, Y. Containing pandemic influenza with antiviral agents. Am J Epidemiol 159, 623-33 (2004).
12. Viboud, C. et al. Risk factors of influenza transmission in households. Br J Gen Pract 54, 684-9 (2004).
13. Principi, N., Esposito, S., Gasparini, R., Marchisio, P. & Crovari, P. Burden of influenza in healthy children and their households. Arch Dis Child 89, 1002-7 (2004).
14. Heymann, A., Chodick, G., Reichman, B., Kokia, E. & Laufer, J. Influence of school closure on the incidence of viral respiratory diseases among children and on health care utilization. Pediatr Infect Dis J 23, 675-7 (2004).
15. Hayden, F. G. et al. Local and systemic cytokine responses during experimental human influenza A virus infection. Relation to symptom formation and host defense. J Clin Invest 101, 643-9 (1998).
16. Elveback, L. R. et al. An influenza simulation model for immunization studies. Am J Epidemiol 103, 152-165 (1976).
17. Longini, I. M., Jr., Koopman, J. S., Haber, M. & Cotsonis, G. A. Statistical inference for infectious diseases. Risk-specific household and community transmission parameters. Am J Epidemiol 128, 845-59 (1988).
18. Fox, J. P., Hall, C. E., Cooney, M. k. & Foy, M. F. Influenza virus infections in Seattle families, 1975-79. I. Study design, methods and the occurrence of infecitons by time and age. Am J Epidemiol 116, 212-27 (1982).
19. Hollingsworth, D., Ferguson, N. M. & Anderson, R. M. How effective were travel advisories in containing the spread of SARS? In preparation (2005).
20. Ferguson, N. M., Fraser, C., Donnelly, C. A., Ghani, A. C. & Anderson, R. M. Public health risk from the avian H5N1 influenza epidemic. Science 304, 968-9 (2004).
16
21. Riley, S. et al. Transmission dynamics of the etiological agent of SARS in Hong Kong: Impact of public health interventions. Science 300, 1961-1966 (2003).
22. World Health Organisation. WHO Inter-country Consultation Influenza A/H5N1 in Humans in Asia. http://www.who.int/csr/disease/avian_influenza/H5N1%20Intercountry%20Assessment%20final.pdf (2005).
23. Ferguson, N. M., Mallett, S., Jackson, H., Roberts, N. & Ward, P. A population-dynamic model for evaluating the potential spread of drug-resistant influenza virus infections during community- based use of antivirals. Journal of Antimicrobial Chemotherapy 51, 977-990 (2003).
24. Herlocher, M. L. et al. Influenza viruses resistant to the antiviral drug oseltamivir: transmission studies in ferrets. J Infect Dis 190, 1627-30 (2004).
25. Oakridge National Laboratory. Landscan global population data. http://www.ornl.gov/sci/gist/landscan (2003).
26. National Statistical Office Thailand. Population Census 2000. http://www.nso.go.th/pop2000/pop_e2000.htm (2000).
27. National Statistical Office Thailand. Population and Labor Statistics. http://www.nso.go.th/eng/stat/subject/subject.htm#cata1 (1999).
28. Ministry of Education Thailand. Class size data for Thai schools. http://www.moe.go.th/ (2005).
29. National Statistical Office Thailand. National Migration Survey. http://www.nso.go.th/eng/stat/migrant/migrant.htm & http://opr.princeton.edu/archive/nmst/ (1994).
30. Chamratrithirong, A. et al. National Migration Survey of Thailand. (Institute for Population and Social Research, Mahidol University, Bangkok, 1995).
31. Ministry of Labour, T. Yearbook of Labour Statistics. http://www1.mol.go.th/download/chap031-2.pdf (2003).
32. Axtell, R. L. Zipf Distribution of U.S. Firm Sizes. Science 293, 1818-1820 (2001). 33. Moser, M. R. et al. An outbreak of influenza aboard a commercial airliner. American
Journal of Epidemiology 110, 1-6 (1979). 34. Fraser, C., Riley, S., Anderson, R. M. & Ferguson, N. M. Factors that make an
infectious disease outbreak controllable. Proc Natl Acad Sci U S A 101, 6146-51 (2004).
35. Hayden, F. G. et al. Use of the oral neuraminidase inhibitor Oseltamivir in experimental human influenza. JAMA 282, 1240-1246 (1999).
36. Carrat, F. et al. Influenza burden of illness: estimates from a national prospective survey of household contacts in France. Arch Intern Med 162, 1842-8 (2002).
37. Cauchemez, S., Carrat, F., Viboud, C., Valleron, A. J. & Boelle, P. Y. A Bayesian MCMC approach to study transmission of influenza: application to household longitudinal data. Statistics in Medicine 23, 3469-3487 (2004).
38. Flahault, A. et al. Modelling the 1985 influenza epidemic in France. Stat Med 7, 1147-55 (1988).
39. Yang, Y. & Longini, I. M. New Methods for the Estimation of Influenza Antiviral Agent Efficacy. in Joint Statistical Meetings (San Francisco, 2003).
17
Figure Legends Figure 1. Data. (a) Modelled population density of Thailand and 100km contiguous zone of neighbouring countries, based on Landscan25 data, plotted on logarithmic scale (light=low density, dark=high density). Inset shows Bangkok in more detail. (b) Age distribution of Thai population in 2003 in 5 year bands and corresponding age distribution of simulated population. (c) as (b) but showing distribution of household sizes. (d) Observed (solid lines) and modelled (dashed lines) distributions of school sizes (blue=elementary, green=secondary, red=mixed). (e) Probability of travelling over a certain distance to work estimated from data and from the simulated population . (f) Weekly excess influenza-related mortality in 1918-19 in GB and corresponding estimates of the reproduction number, R, calculated assuming gT =2.6. (g) Viral shedding data for experimental influenza infection35 compared with the modelled profile of infectiousness over time. Note that the infectiousness profile was not fitted to shedding data. See Methods and Supplementary Information for more details. Figure 2. Expected pattern of spread of an uncontrolled epidemic. (a) Sequence of snapshots of the epidemic showing extent of spread of a single simulation of a 0R =1.5 epidemic. Red indicates presence of infectives, green the density of people who have recovered from infection or died. (b) Daily incidence of infection over time in the absence of controls for 0R =1.5. Thick blue line represents average for realisations resulting in a large epidemic, grey shading represents 95% envelope of incidence timeseries. Multiple coloured thin lines show a sample of realisations, illustrating large degree of stochastic variability. (c) Root Mean Square (RMS) distance from seed infective of all individuals infected since the start of the epidemic as a function of time. Thick blue line represents average distance for realisations resulting in a large epidemic, grey shading represents 95% envelope. (d) Proportion of the population infected by age averaged across realisations giving large epidemics, for R0 = 1.5. The infection attack rate is 33% for 0R =1.5, and 50% for 0R =1.8. (e) Distribution of number of secondary cases per primary case during the exponential growth phase of a
0R =1.5 epidemic. Between 50 and 1000 realisations were used to calculate all averages (see Supplementary information). Figure 3. Prophylaxis strategies. We assume 90% of clinical cases (=45% of infections) are detected. Social targeting assumes prophylaxis of 90% of household members and 90% of pupils or colleagues in 90% of the schools or workplaces of detected cases. Geographic targeting assumes 90% of people within 5, 10 or 15km of a detected case are also prophylaxed. (a) Probability of eliminating an otherwise large epidemic using social plus geographic targeting, as a function of 0R of the new strain and the radius of prophylaxis. Results assume policy initiation after 20 cases detected and a 2 day delay from case detection to prophylaxis. Error bars show exact 95% confidence limits. (b) as (a) but showing average numbers of drug courses required for containment of an otherwise large outbreak. (c) Map (of 150x150km square of northern Thailand) showing extent of spread during one contained
0R =1.8 epidemic assuming 10k radial prophylaxis and other parameters as (a). Blue represents treated areas. (d) as (a), but varying the delay (from 0-4 days) from case detection to prophylaxis for the 5km radius policy. (e-f) as (a-b) but for drug-sparing policies which target only the nearest 10, 20, 30 or 50 thousand people within 10km of a detected case. . Figure 4. Social distance measures. (a-b) as Fig 3(a-b), but showing impact of drug-sparing prophylaxis (50,000 courses per case, as Fig 3e) combined with: no social distance measures (as Fig 3); closure of 90% of schools and 50% of workplaces within 5km of a detected case for 21 days; 80% ‘area quarantine’ (i.e. reduction by 80% of travel in and out of a zone defined by merging 5km rings around all detected cases) for 21 days; combination of school/workplace closure and area quarantine. (c) as (a) but showing effect of limiting availability of antiviral drugs to 1, 3 or 5 million courses on the effectiveness of the combined area quarantine and 5km radial prophylaxis policy. (d) as (c) but for drug-sparing geographic prophylaxis (50,000 courses per case) plus area quarantine. (e) Case incidence over time without pandemic control and with the 3 million course policy of (d), showing the
18
approximate 1 month delay achieved even when containment is unsuccessful ( 0R =1.9). (f) as (a), but showing the reduction in policy effectiveness seen if the combined school/workplace closure and drug-sparing prophylaxis policy is restricted to Thailand alone.
0
0.5
1
1.5
2
0 2 4 6 8 10
Days post infection
Infe
ctio
usne
ss0
1000
2000
3000
4000
5000
Vira
l loa
d (d
ata)
0.001
0.01
0.1
1
1 10 100 1000Distance (km)
Cum
ulat
ive
prob
abili
ty
Data
Model
0
0.5
1
1.5
2
Jun-18 Sep-18 Dec-18 Apr-19Date
R
0
2000
4000
6000
8000
Wee
kly
deat
hs
RDeaths
0
0.05
0.1
0.15
0.2
0.25
0.3
1 2 3 4 5 6 7 8 9 1
Household size
Pro
porti
on o
f ho
useh
olds model
data
0
0.02
0.04
0.06
0.08
0.1
0 10 20 30 40 50 60 70 80
Age band (years)
Pro
porti
on o
f po
pula
tion model
data
00.020.040.060.080.10.120.140.16
0 500 1000 1500 2000
School size
Pro
babi
lity
a b
c
d
e
f
g
0
200
400
600
800
1000
0 50 100 150 200 250Day
Dist
ance
(km
)
0
0.1
0.2
0.3
0.4
0 1 2 3 4 5 6 7 8 910Secondary infections
Prop
ortio
n
0100000
200000300000400000
500000600000
0 50 100 150 200 250Day
Infe
ctio
ns
0%10%20%30%
40%50%60%
0 10 20 30 40 50 60 70 80Age group (years)
Prop
ortio
n in
fect
ed
a
30 60 90 120 150 180 240
b c ed
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Prob
abili
ty o
f el
imin
atio
n
10,00020,00030,00050,000
0E+05E+51E+62E+62E+63E+63E+6
1.1 1.3 1.5 1.7R 0
Aver
age
num
ber
of c
ours
es
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Prob
abili
ty o
f el
imin
atio
n
01234
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Prob
abili
ty o
f el
imin
atio
n
0km5km10km15km
0E+0
1E+6
2E+63E+6
4E+6
5E+6
1.1 1.3 1.5 1.7R 0
Aver
age
num
ber
of c
ours
es
a
c
e
b
d
f
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Prob
abili
ty o
f el
imin
atio
nNonePlacesMovementsPlaces+Moves
0E+0
5E+5
1E+6
2E+6
2E+6
1.1 1.3 1.5 1.7 1.9R 0
Aver
age
num
ber
of c
ours
es
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Prob
abili
ty o
f el
imin
atio
n
ThailandonlyTrans-national
110
1001000
10000100000
1000000
0 20 40 60 80 100Day
Daily
cas
es
No control
With control
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Prob
abili
ty o
f el
imin
atio
n
1m3m5mUnlimited
50%
60%70%
80%90%
100%
1.1 1.3 1.5 1.7 1.9R 0
Pro
babi
lity
of
elim
inat
ion
1m3m5mUnlimited
a b
c d
e f